
Event beta: realized loading of asset returns on event-probability news
Source:R/beta.R
event_beta.RdRegresses asset returns on the innovations of a traded event
probability, $$r_t = \alpha + b\, \Delta q_t + \gamma' c_t + u_t,$$
with Newey-West (Bartlett/HAC) standard errors. To first order the
model implies \(r_t \approx \Delta\eta\,\Delta q_t\), so the slope
b is the returns-based estimate of the event exposure
\(\widehat{\Delta\eta}\).
Usage
event_beta(asset, x, deta = NULL, controls = NULL, lags = NULL)
# S3 method for class 'event_beta'
print(x, ...)Arguments
- asset
A
data.framewith a time column and either a return column (ret) or a price column (adjusted,close, orprice; log returns are computed).- x
An
event_pricesobject (or coercible); its probability innovations \(\Delta q_t\) are the regressor. The probability level is taken at each asset date (the last observation on or before it) and differenced between consecutive asset dates, so probability moves over weekends and holidays enter the next trading day's return. Collapse intraday prices to a daily snapshot first (seepm_daily()).- deta
Optional externally measured event exposure \(\Delta\eta\); enables the \(\beta = 1\) test.
- controls
Optional
data.framewith a time column and control variables (e.g. market returns), matched on date.- lags
Newey-West lag order; default \(\lfloor 4 (n/100)^{2/9} \rfloor\). Use
lags = 0for heteroskedasticity-robust (HC0) errors.- ...
Unused (for the
printmethod).
Value
An object of class event_beta: a list with
- coefficients
tibble of terms, estimates, HAC standard errors, t-statistics, and p-values.
- deta_hat, deta_se
the slope on \(\Delta q\) and its SE.
- eta1, eta2
model-implied levels (see Details).
- beta, beta_se, beta_z, beta_p
(only with
deta) the loading \(b/\Delta\eta\) and the Wald test of \(\beta = 1\).- r2, n, lags
regression diagnostics;
r2is the realized variance share of event news over the sample.
Details
What is (and is not) identified. Returns identify only the
spread \(\Delta\eta = \eta_1 - \eta_2\), not the
outcome-conditional levels \(\eta_1, \eta_2\) separately (those
require event-spanning option smiles). Consequently the loading test
\(\beta = 1\) is only meaningful against an externally measured
exposure: supply deta (e.g. option-implied, or from an independent
sample) and the function reports \(\beta = b / \Delta\eta\) with a
Wald test of \(\beta = 1\). Without deta, the regression is
exactly identified and only \(\widehat{\Delta\eta}\) is reported.
Under the pricing-measure adding-up constraint
\(q\,\eta_1 + (1-q)\,\eta_2 = 1\), point estimates of the levels can
be backed out as \(\eta_1 = 1 + (1-q)\Delta\eta\) and
\(\eta_2 = 1 - q\Delta\eta\); these are model-implied, not
independently identified, and are returned for convenience
(evaluated at the sample-average q).
See also
ec_relevance() for the pricing-relevance screen.
Examples
data(djt2024)
data(polymarket2024)
ep <- pm_daily(as_event_prices(polymarket2024))
event_beta(djt2024, ep)
#> -- Event-beta regression (Newey-West, 4 lags)
#> Event exposure deta_hat = 1.4669 (se 0.3728, t = 3.93), n = 125, R^2 = 0.106
#> Model-implied levels at mean q = 0.554: eta1 = 1.6549, eta2 = 0.1881
#> (No external deta supplied: levels/loading test not identified from returns alone.)